Reflections on Google I/O

Last post, I wrote about how the threat posed by ChatGPT to Google's search dominance was greatly overstated, and that Google would likely continue to thrive amidst the generative AI revolution. At the end of April, I tweeted a prediction that Google's stock would pop during Google I/O. The thinking behind that prediction was that Google would not repeat the mistake of their initial botched Bard demo, which seemed bizarrely rushed and reactionary, and ultimately wiped out $100B of Alphabet's market cap in a day. No, Google has well-developed organizational muscles around deploying AI at scale and too much to lose to again miss an opportunity to convey their readiness to meet the current AI moment.

Google I/O was indeed the show of AI force I thought it would be, featuring demos of generative AI in Search, Workspace (Gmail, Docs, Slides, etc), Android, as well as some demos of Bard (including the announcement that the waitlist for Bard had been removed and was free to use across 180 countries). As of this writing, Google's stock is up 14.2% since I/O and just had it's best month in 3 years.

It's always nice to have a prediction proven correct, especially when that prediction is about the upward trend of a stock that you own, but Google I/O felt distinctly, and surprisingly, underwhelming. The main thing that stuck out to me as that AI is clearly an evolutionary, not a revolutionary, new technology in the Big Tech toolkit. The much advertised "Help me write" feature in Gmail, for example, is a clear continuation of earlier assisted writing features.

In Search, as predicted, the experience for queries with commercial intent (Google's moneymaker) appears to be relatively unchanged. Ads are still important, so you will still see them and be encouraged to click them.

And in Workspace, there weren't really any capabilities demoed you couldn't already get from already available tools such as Midjourney or ChatGPT.

The following pair of screenshots shows some image generation capabilities embedded into Slides.

‍

And the next screenshot shows how generative AI functionality is embedded into Docs.

In a vacuum, these capabilities are impressive. But given the state of the art of many freely available tools, these register more as minor add-ons to existing products as opposed to significant changes or new product categories altogether.
‍

All in all, the message, rather explicit in the form of slides packed with icons, from the keynote was that Alphabet builds a lot of wide reaching products, and they are going to ram generative AI into as many of them as they can.

‍

If generative AI infused products are the new standard for consumer and business software, then Alphabet convincingly showed that they will be able to meet that standard. I can see why a Wall Street analyst would find that message encouraging. The plain and sad truth of things is that all Alphabet really has to do is meet the standard, they do not have to exceed it or impress and delight their users. Alphabet's products are invariably buoyed so much by their scale, reach, portfolio, and switching costs that as long as they don't build outright inferior products, they will continue to survive and thrive. And Alphabet demonstrated they have the organizational agility and self-awareness to continue to make use of those advantages.

Coincidentally, the month of May also included an announcement that Neeva, the privacy focused search engine billed as a Google competitor, would be shutting down it's consumer search product. In their announcement, they noted:

"But throughout this journey, we’ve discovered that it is one thing to build a search engine, and an entirely different thing to convince regular users of the need to switch to a better choice. From the unnecessary friction required to change default search settings, to the challenges in helping people understand the difference between a search engine and a browser, acquiring users has been really hard. Contrary to popular belief, convincing users to pay for a better experience was actually a less difficult problem compared to getting them to try a new search engine in the first place."

As mentioned last time, these switching costs are Google's true moat.

Beyond mere underwhelm, though, I started to get a mild case of AI doomerism. From what I could glean, the best case scenario of Alphabet's vision of AI-infused products is that it will be a mild labor-savor for current users, and I don't want to diminish that. But I'm not seeing any evidence that Alphabet has actual user satisfaction in mind. Sundar's keynote feature an eye popping but ultimately useless metric around how many times AI features have been used. How does that translate into anything I, as a user, would care about? Does it save me time? Does it help me produce higher quality docs/slides/etc?

At worst, though, I worry that these "AI features" are going to make us more passive, more inept, and less creative. AI-assisted writing, in the sense that you click buttons or write short prompts to write much larger/longer documents, kind of seems like robo-assisted weightlifting. You lose the benefit by making it too easy. Obviously, making the tedious stuff easier is a good thing. But given the overall usage metrics Alphabet appears to be optimizing for, there will not be much of a distinction between the tedious stuff and the Real Writing. The point will be to get people to use the AI, even if they lose their own aptitude in the process.

I'm certainly not the first person to arrive at these concerns. Paul Graham, Silicon Valley's favorite essayist, warned that when "you lose the ability to write, you also lose some of your ability to think." Jaron Lanier, in his book "Dawn of the New Everything" (published in 2017, before the LLM hype), advances a similar argument quite thoughtfully:

"...we make ourselves dumb to make computers look smart all the time.

Consider Netflix. The company claims that its smart algorithm gets to know you and then recommends movies. The company even offered a million-dollar prize for ideas to make the algorithm smarter.

The thing about Netflix, though, is that it doesn't offer a comprehensive catalog, especially of recent, hot releases. If you think of any particular movie, it might not be available for streaming. The recommendation engine is a magician's misdirection, distracting you from the fact that not everything is available.

So is the algorithm intelligent, or are people somewhat blind and silly in order to make the algorithm seem intelligent?"

We are seeing this same dynamic play out in AI-assisted writing. We are forgetting the universe of options we have available to us when writing and instead being impressed by a bot that can produce plausible but stock responses. But writing is a much more core cognitive skill than selecting movies, and as Paul Graham warns, there will be consequences.

It's really not all doom and gloom though. There is a lot in AI to be excited about despite the fact that it seems big tech companies are not going to be the purveyors of a real AI revolution. The exciting innovations, the ones that are aiming to make AI can unlock new computing modalities altogether, rather than just as plugins to existing products, and doing so in ways that will actually help humans, seems to be coming from startups...a few that I'm excited about are inflection.ai, adept.ai, and rewind.ai. All seem to be circling, in different ways, the idea of a truly personalized, always-on/cross-application AI, a true digital assistant that augments human intelligence. To unpack this a bit:

• Truly Personalized - I want an AI that deeply understands my behaviors and goals. I don't want an AI that just spits back uncontroversially safe or generally pretty good suggestions. The generic AI is the one that seems smart by virtue of me making myself dumb. Instead, I want an AI that makes me smarter in its presence the same way a tutor that really understands my learning style makes me smarter. inflection.ai is leaning into this concept so much that their first product is called PI, for Personal Intelligence.
• Always-On/Cross-Application - The more the AI knows about me, and about what I do, the more it can help. In my view, that means the killer AI digital assistant of the future cannot be relegated to a browser window. adept.ai and rewind.ai are both leaning into this idea very intentionally. Throughout the course of my day, I am using probably eight different programs - my programming editor, outlook, slack, zoom, my Evernote, my browser, and various developer tools. A ubiquitous and truly helpful digital assistant will understand my usage patterns across all these applications, and in doing so, understand and help me far more than it could if it was just a website.

I'm excited to follow the progress of these and similar companies, and eager to use their products once they become available.

‍